Measurement of linear accelerator spectra, reconstructed from percentage depth dose curves by neural networks

Purpose Linear accelerator (linac) spectra, used to improve the accuracy of dose calculation and to produce a complete description of beam quality, among other aspects, are relevant in radiotherapy and linear accelerator physics. Methods In this work we apply neural networks in solving an ill-conditioned system of linear equations, to indirectly measure the linear accelerator spectra via the percentage depth dose curves. The photon beam spectra are related to radiation doses through a Fredholm integral equation. To address our problem we measured the percentage depth dose curve in water and solved a discretized Fredholm equation using artificial neural network. After analysing the typology of our physical problem we selected a MultiLayer Perceptron Neural Network and designed the most suitable neural network architecture. Results The reconstructed spectra were compared with spectra from three linacs, two of them obtained by us through simulations and the third produced by another author, achieving a concordance between 92 % and 96 %. Conclusions Therefore, the spectrum of any accelerator can be quickly and easily reconstructed from measured percent depth dose curves, applying a trained artificial neural network.